Machine Learning · arXiv:2611.20064 v1 · submitted 15 Nov 2026 · 23 references
Decode Early, Verify Late: Streaming Offline Speech Recognizers Without Accuracy Loss
The paper says
“Voice agents need accurate transcripts soon after a user stops speaking. We propose Decode Early, Verify Late (DEVL), a training-free framework for streaming inference with offline ASR models.”
more of the abstract
The server keeps partial transcripts as drafts and verifies the latest against the complete audio in one forward pass at turn end. Under stated input and decoding conditions, the final equals the model's greedy offline output in exact arithmetic. We adapt DEVL to four model families. In our five-recognizer comparison, 95-100% of final transcripts are byte-identical to the offline transcripts, WER stays within 0.13 points of offline, and single-stream time to final falls 1.5-7.3x. Serving Qwen3-ASR-1.7B on one A100, DEVL returns finals a median 22-26 ms after turn end at up to 64 streams, 2.7-3.9x sooner than the official streaming mode. It is the only system we tested that meets a 100 ms p99 target, up to 80 streams on Pipecat-STT and 64 on WildASR.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 17 Nov 2026 · read the full text (34,891 characters) · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A practical method for converting offline ASR models to streaming without accuracy loss, with thorough latency measurements and a clear correctness argument. This line is wrong
15 minor · 2 open questions
How often objections from reviews like this one are right: not yet measured
The paper, first character to last
The reviewer was given the whole text, 34,891 characters. Not checked: figures, code, what the 23 references say.
15 minor objections
The paper says, at §2, Algorithm 1
Pith says: minor objection 1 of 15
The algorithm pseudocode uses notation like $d_{1:\max(k-1,0)}$ which is slightly confusing when $k=0$. Consider clarifying the edge case for empty drafts.
The paper says, at §3.1, Table 1 caption
Pith says: minor objection 2 of 15
The caption for Table 1 is very long and contains multiple footnotes and explanations. Consider moving some of the detailed notes to the main text or a separate footnote to improve readability.
show 13 more
The paper says, at §4.1, Table 2 (exact-rows.tex not visible)
Pith says: minor objection 3 of 15
The exactness results for Qwen3-ASR by load are presented in a table that is not visible in the supplied text (data-cap/exact-rows.tex). The paper should include the actual numbers in the main text or ensure the table is self-contained.
The paper says, at §4.3, Table 3 (other-load-rows.tex not visible)
Pith says: minor objection 4 of 15
Similar to the previous comment, the capacity results for other recognizers are in an external file that is not visible. The paper should include these numbers in the main text or appendix.
The paper says, at §5, Limitations
Pith says: minor objection 5 of 15
The paper states that 'Our benchmarks are English, and Pipecat-STT's references are machine-generated.' This is an important limitation that should be mentioned earlier, perhaps in the abstract or introduction, to set expectations for readers.
The paper says, at §2, Algorithm 1, line 'at most three repair rounds'
Pith says: minor objection 6 of 15
The algorithm mentions 'at most three repair rounds' but does not explain why three was chosen or how the number of rounds affects correctness or latency. A brief justification would be helpful.
The paper says, at §2, Algorithm 1, 'epoch' handling
Pith says: minor objection 7 of 15
The algorithm mentions epoch cuts for utterances longer than 10 s, but the correctness guarantee explicitly excludes epoch-cut finals. The paper should clarify how epoch cuts interact with the verification process and whether the same invariants hold within each epoch.
The paper says, at §4.1, Table 1, Whisper-Streaming footnote
Pith says: minor objection 8 of 15
The paper attributes most of Whisper-Streaming's accuracy gap to 'default decoding settings' but does not provide a direct comparison with greedy decoding on the same backend. The footnote is vague and could be misinterpreted. Consider providing the actual WER with greedy decoding on the Whisper-Streaming backend.
The paper says, at §4.2, Figure 2 (not visible)
Pith says: minor objection 9 of 15
The load testing figure is not visible in the supplied text. The caption and description are clear, but the actual data points and error bars cannot be assessed. This is a limitation of the review process, not the paper, but the paper should ensure the figure is self-explanatory.
The paper says, at §4.2, 'Letting 16 or 32 streams run checkpoints at once keeps 95% or more checked'
Pith says: minor objection 10 of 15
This sentence is somewhat unclear: does it mean that the scheduler allows 16 or 32 streams to run checkpoints concurrently? The paper previously stated 'at most four streams active except when epoch cuts require more.' There is a potential inconsistency that should be clarified.
The paper says, at §4.3, 'VibeVoice-ASR loops on two clips'
Pith says: minor objection 11 of 15
The paper mentions that VibeVoice-ASR loops on two clips and that excluding them gives 11.3-12.8% WER. This is an important detail that affects the interpretation of the results. Consider reporting both with and without the problematic clips in the main table.
The paper says, at §2, 'Cached representations are recomputed if new audio changes them'
Pith says: minor objection 12 of 15
The paper states that cached representations are recomputed if new audio changes them, but it is not clear what triggers this recomputation. Does the system detect changes in input normalization automatically? A brief explanation would be helpful.
The paper says, at §2, 'The scheduler runs pending finals before checkpoints'
Pith says: minor objection 13 of 15
The paper mentions that the scheduler prioritizes finals over checkpoints, but does not describe the scheduling algorithm in detail. How are deadlines or priorities assigned? This is relevant for reproducibility.
The paper says, at §3.1, 'For VibeVoice-ASR, this is the joined JSON \texttt{Content} text'
Pith says: minor objection 14 of 15
The paper defines 'identical' as byte-identical final transcript text, but for VibeVoice-ASR it uses the joined JSON Content text, excluding timestamps and speaker metadata. This is a reasonable choice, but it should be stated more clearly in the main text rather than in a parenthetical remark.
The paper says, at §4.1, 'Where we recorded logits for \sys{}'s differing finals, the first divergence was at a near-tie between the top…
Pith says: minor objection 15 of 15
This is an interesting observation, but it is not quantified. How many differing finals were analyzed? What fraction had near-ties? Providing numbers would strengthen the claim.
Open questions (2)
Pith raised these and could not confirm them against the text of the paper it read. They are questions for the authors, not findings.
- §4.2, Figure 2 (not visible), §4.3 How many runs were performed at each load level for the capacity measurements, and what is the variability of the capacity estimates?
- §4.3, Table 3 (other-load-rows.tex not visible) What are the full capacity numbers for the other recognizers (Whisper large-v3, VibeVoice-ASR, Nemotron) at each p99 TTFS target, as would appear in the table referenced by \inputrows{data-cap/other-load-rows.tex}?
What would settle it
Run DEVL on a model with stochastic decoding (e.g., beam search with temperature > 0) and check whether the final transcript still matches the offline output. The paper's guarantee explicitly requires greedy decoding, so a mismatch would confirm the assumption's necessity rather than refute the paper, but it would show the framework's limited scope.
from Pith's summary of the paper
How the review went
5 passes by deepseek-v4-flash · 17 Nov 2026
- Reader holds up, with conditions
- Skeptic concern withdrawn: Pith's reviewer misread the paper
- Referee 15 minor
- Check against the paper kept 0 of 5 objections
- Editor wrote the verdict above
Not run on this paper: the simulated authors' defence.
Pith's passes raised 5 objections. A check against the paper's text kept 0, withdrew 3 (3 Pith's reviewer misread the paper) and left 2 as questions.
The review in full
The review in brief, machine-written
This paper proposes Decode Early, Verify Late (DEVL), a training-free framework that reuses partial transcripts as drafts and verifies them against the complete audio at turn end, so that an offline ASR model can serve streaming requests while preserving its greedy decoding output. The core idea is well-motivated: the most accurate open-weight ASR models are offline, and existing streaming options either sacrifice accuracy or add latency. DEVL adapts speculative decoding to the ASR setting, with careful attention to audio representations, decoder state, and prompt invariants.
The paper does several things well. It states explicit conditions under which the final transcript equals the greedy offline output in exact arithmetic, and it provides a correctness sketch. The empirical evaluation is thorough: five recognizers, two datasets, load testing up to 128 streams, and comparisons against multiple baselines. The latency results are striking—Qwen3-ASR meets a 100 ms p99 TTFS target at 64–80 concurrent streams on one A100, which no other system in the comparison achieves. The adaptations to four model families (Qwen3-ASR, Whisper, VibeVoice-ASR, Nemotron) show that the method generalizes beyond a single architecture.
The soft spots are real but acknowledged. The exactness guarantee is qualified by batch-dependent GPU rounding and by epoch cuts for long utterances; the paper reports that batching alone changes up to 1.9% of finals at 128 streams. The comparison against Whisper-Streaming uses default decoding settings that account for most of the accuracy gap, which the paper notes but does not correct. No code is provided, which limits reproducibility. The benchmarks are English-only, and Pipecat-STT's references are machine-generated. The method requires specific model properties (greedy decoding, deterministic front end, restorable decoder state) that may not hold for all ASR models.
This paper is for practitioners deploying offline ASR models in latency-sensitive voice agent applications. It deserves a serious referee: the method is clearly described, the correctness conditions are stated, and the empirical evaluation is substantial. The main revisions to ask for are releasing code, addressing the Whisper-Streaming comparison with matched decoding settings, and clarifying the variability of the capacity estimates. I would accept this paper for peer review.
The stress test's concern
Withdrawn: Pith's reviewer misread the paper.
The paper says
“Under stated input and decoding conditions, the final equals the model's greedy offline output in exact arithmetic.”
The referee's report
Not published. Pith's referee wrote this assessment before its objections were checked against the paper, and 4 of its 4 major objections did not survive that check.
Pith's reading of the paper (a machine's summary, not the authors')
The paper proposes DEVL, a training-free framework that converts offline ASR models into streaming recognizers without sacrificing accuracy. It works by keeping partial transcripts as drafts during speech and verifying the latest draft against the complete audio in a single forward pass at turn end. Under stated conditions (greedy decoding, deterministic front end, restorable decoder state), the final transcript equals the model's greedy offline output in exact arithmetic. Empirically, across five recognizers, 95-100% of finals are byte-identical to offline transcripts, WER stays within 0.13 points, and single-stream time to final drops 1.5-7.3x. The paper argues this makes offline-quality streaming practical for voice agents.
What carries the argument. The central mechanism is the verification of a draft partial transcript against the complete audio in a single forward pass at turn end, analogous to speculative decoding. The paper defines conditions (greedy decoding, deterministic front end, restorable decoder state) under which this verification guarantees the final equals the greedy offline output in exact arithmetic. The key objects are the draft (a partial transcript kept from a checkpoint), the KV cache (which must be restored to the accepted prefix), and the verification pass that checks the draft against the complete audio and then decodes only after the longest correct prefix.
Core claim. The paper claims that by reusing partial transcripts as drafts and verifying them against the complete audio at turn end, an offline ASR model can produce streaming finals that are identical to its greedy offline output under exact arithmetic, provided the decoder state can be restored to an accepted prefix and the audio representations, prompt, token positions, and retained decoder state equal those of an offline pass. This is demonstrated across five recognizers (Qwen3-ASR, Whisper large-v3-turbo, Whisper large-v3, VibeVoice-ASR, NVIDIA-Nemotron-3.5-Transcribe) with 95-100% byte-identical transcripts and WER within 0.13 points of offline. The paper also reports that serving Qwen3-ASR-1.7B on one A100 with DEVL returns finals a median 22-26 ms after turn end at up to 64 streams, 2.7-3.9x sooner than the official streaming mode, and is the only system tested that meets a 100 ms p99 TTFS target at up to 80 streams on Pipecat-STT and 64 on WildASR.
The premise it rests on. The correctness guarantee assumes that the decoder state can be restored to an accepted prefix and that the audio representations, prompt, token positions, and retained decoder state equal those of an offline pass over the same audio, which requires greedy decoding, a deterministic front end, and restorable decoder state.
If the paper is right (Pith's inferences, not the authors')
Voice agents can use the most accurate open-weight ASR models (which are offline) in streaming mode without accuracy loss, closing the gap between offline and streaming WER.
The time to final transcript after a user stops speaking can be reduced to tens of milliseconds (22-26 ms median for Qwen3-ASR on one A100), enabling more responsive voice interfaces.
The framework is training-free and adapts to multiple model families (Qwen3-ASR, Whisper, VibeVoice-ASR, Nemotron) with different encoder architectures, suggesting broad applicability.
At low to moderate concurrency (up to 64 streams on one A100), DEVL meets strict latency targets (100 ms p99) that other streaming approaches miss, potentially reducing the number of GPUs needed for a given throughput.
The paper's exactness guarantee (final equals greedy offline output in exact arithmetic) is limited to models with greedy decoding, deterministic front ends, and restorable decoder state. Models using beam search or stochastic decoding would need a different verification strategy, and the paper does not address those cases.
The empirical results show that batch-dependent GPU rounding changes some finals even without drafts (up to 1.9% at 128 streams for Qwen3-ASR), so the exactness guarantee is a mathematical ideal that real hardware approximates. The paper acknowledges this and suggests batch-invariant kernels as a fix.
The paper's focus on English benchmarks and machine-generated references (Pipecat-STT) leaves open how DEVL performs on other languages and on human-transcribed ground truth. The WildASR dataset includes accented English but is still English-only.
The paper reports that checkpoints consume GPU time and help slower models mainly at low load. For models like VibeVoice-ASR and Nemotron, drafts stop helping at higher concurrency (32-64 streams), suggesting the framework's benefit is most pronounced for fast models or low load.
Check this review
What was read
The passes were given all 34,891 characters of the paper's LaTeX source.
Fingerprint of that text (SHA-256): 4a943482278c24d3c914be0d00765ae2835fdc4804bd02d0fb431785f886e2ac
Pith does not keep that text. Rebuilt from the paper's source or PDF by the same code, it has this fingerprint if it is the same text. From a copy of Pith's code:
PYTHONPATH=. python scripts/review_input_check.py 2611.20064
The passes, in the order they ran
- Reader deepseek-v4-flash · medium effort · $0.0024 · instructions of the Reader, fingerprint f9ead7300382…
- Skeptic deepseek-v4-flash · medium effort · $0.0015 · instructions of the Skeptic, fingerprint 8f88c0f61488…
- Referee deepseek-v4-flash · medium effort · $0.0019 · instructions of the Referee, fingerprint a852d99aea0d…
- Check against the paper deepseek-v4-flash · medium effort · $0.0020 · instructions of the Check against the paper, fingerprint 6c7a6883d551…
- Editor deepseek-v4-flash · medium effort · $0.0015 · instructions of the Editor, fingerprint e7211f6a3d27…
- Circularity check deepseek-v4-flash · medium effort · $0.0017 · instructions of the Circularity check, fingerprint e9946e39d67b…
- Summary for the page deepseek-v4-flash · medium effort · temperature 0.2 · $0.0018 · instructions of the Summary for the page, fingerprint e4f56bf568e7…
The passes cost $0.013 in model calls together.
Version stamped on this review: pith-pipeline@v1.4.1-daily-deepseek-or-medium+deepseek/deepseek-v4-flash+p.d60f54226f00
Each link opens a pass's instructions as plain text, the exact bytes its fingerprint was taken over. Every stored set of instructions.
curl -s https://pith.science/methodology/prompts/f9ead730038216984cfb56d7cb34332bdf691828296a1de4245f3d6667d4bf56 | sha256sum
prints the fingerprint of the reader's instructions. The same works for every pass.
The receipt
The review's fingerprint was taken on 17 Nov 2026. The review you are reading still matches it.
Fingerprint of the review (SHA-256): 557230fde0e3e834bbbeaa07e3425954a679f597bc2a195bede37dad509c43df
curl -s 'https://pith.science/api/v1/verdicts/2611.20064?form=canonical' | sha256sum
prints the fingerprint of the review as it stands now. The rule, in four steps.
A matching fingerprint shows the review has not changed since that date. It does not show that the review is right.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, from the paper's LaTeX source.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Voice agents need accurate transcripts soon after a user stops speaking. We propose Decode Early, Verify Late (DEVL), a training-free framework for streaming inference with offline ASR models. The server keeps partial transcripts as drafts and verifies the latest against the complete audio in one forward pass at turn end. Under stated input and decoding conditions, the final equals the model's greedy offline output in exact arithmetic. We adapt DEVL to four model families. In our five-recognizer comparison, 95-100% of final transcripts are byte-identical to the offline transcripts, WER stays within 0.13 points of offline, and single-stream time to final falls 1.5-7.3x. Serving Qwen3-ASR-1.7B on one A100, DEVL returns finals a median 22-26 ms after turn end at up to 64 streams, 2.7-3.9x sooner than the official streaming mode. It is the only system we tested that meets a 100 ms p99 target, up to 80 streams on Pipecat-STT and 64 on WildASR.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-11-17
The authors' main claim, as the classifier picked it out 0.96
“In our five-recognizer comparison, 95-100% of final transcripts are byte-identical to the offline transcripts, WER stays within 0.13 points of offline, and single-stream time to final falls 1.5-7.3x.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.13 · states limitation 0.77 · human participants 0.06 · controlled comparison 0.92 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.93 · mentions code or data availability 0.04
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Voice agents need accurate transcripts soon after a user stops speaking. | background 1.00 | 0.06 | 0.02 | asserted 0.98 |
| We propose Decode Early, Verify Late (DEVL), a training-free framework for streaming inference with offline ASR models. | contribution 0.98 | 0.88 | 0.12 | asserted 0.99 |
| The server keeps partial transcripts as drafts and verifies the latest against the complete audio in one forward pass at turn end. | method 0.99 | 0.73 | 0.09 | asserted 1.00 |
| Under stated input and decoding conditions, the final equals the model's greedy offline output in exact arithmetic. | result 0.73 | 0.77 | 0.22 | asserted 1.00 |
| We adapt DEVL to four model families. | contribution 0.51 | 0.92 | 0.03 | asserted 0.99 |
| In our five-recognizer comparison, 95-100% of final transcripts are byte-identical to the offline transcripts, WER stays within 0.13 points of offline, and single-stream time to final falls 1.5-7.3x. | result 1.00 | 0.96 | 0.48 | asserted 1.00 |
| Serving Qwen3-ASR-1.7B on one A100, DEVL returns finals a median 22-26 ms after turn end at up to 64 streams, 2.7-3.9x sooner than the official streaming mode. | result 1.00 | 0.94 | 0.13 | asserted 1.00 |
| It is the only system we tested that meets a 100 ms p99 target, up to 80 streams on Pipecat-STT and 64 on WildASR. | result 0.98 | 0.92 | 0.19 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-17: 1 request, 6,264 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (2)
Figures are the authors', reproduced from the paper.
References (23)
- [1] Y.Bian,Y.Zhang,Y.Wang,L.Zhang,R.He,Z.Wang,etal. Decodebeforeyouhear: Speculativelistening for chunk-interleaved full-duplex speech language models, 2026. arXiv:2611.06982 on PithPith's reviewarXiv
- [2] Coval. Coval benchmarks.https://github.com/coval-ai/benchmarks, 2026. stt-v3 and WildASR manifests v1.0.0 at commit db1e3d4. Accessed November 2026 on Pith
- [3] H. He and Thinking Machines Lab. Defeating nondeterminism in LLM inference. https: //thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/, Sept. 2025. Thinking Machines Lab: Connectionism on Pith
- [4] W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, et al. Efficient memory management for large language model serving with PagedAttention. InProc. 29th Symposium on Operating Systems Principles (SOSP), pages 611–626. ACM, 2023. 5 on Pith
- [5] Y. Leviathan, M. Kalman, and Y. Matias. Fast inference from transformers via speculative decoding. In Proc. International Conference on Machine Learning (ICML), volume 202 ofProceedings of Machine Learning Research, pages 19274–19286. PMLR, 2023 on Pith
- [6] Z.-W. Liou, H. Nover, A. Naik, A. Skliar, I. Shovkun, I. Shahaf, and T. Dao. ReplaySSM: Cache SSM inputs, not state, 2026. arXiv:2610.13778 on PithPith's reviewarXiv
- [7] D. Liu, G. Spanakis, and J. Niehues. Low-latency sequence-to-sequence speech recognition and translation by partial hypothesis selection. InProc. Interspeech, pages 3620–3624, 2020 on Pith
- [8] D. Macháček, R. Dabre, and O. Bojar. Turning Whisper into real-time transcription system. InProc. 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics: System Demo... on Pith
- [9] Nari Labs. Nari labs leads coval’s voice ai benchmarks. https://narilabs.com/blog/ nari-labs-leads-coval-voice-ai-benchmarks/ , 2026. Coval speech-to-text benchmark, 1-day view on 14 September 2026. Accessed November 2026 on Pith
- [10] NVIDIA. nemotron-3.5-asr-streaming-0.6b. https://huggingface.co/nvidia/nemotron-3. 5-asr-streaming-0.6b, 2026. Model card, revision ea30d66. Accessed November 2026 on Pith
- [11] NVIDIA. NVIDIA-Nemotron-3.5-Transcribe-30B-A4B. https://huggingface.co/nvidia/ NVIDIA-Nemotron-3.5-Transcribe-30B-A4B ,2026. Modelcard,revisionc176676.AccessedNovem- ber 2026 on Pith
- [12] Z. Peng, J. Yu, Y. Chang, Z. Wang, L. Dong, Y. Hao, et al. VibeVoice-ASR technical report, 2026. arXiv:2601.18184 on PitharXiv
- [13] Pipecat. stt-benchmark-data: Speech-to-text benchmark data.https://huggingface.co/datasets/ pipecat-ai/stt-benchmark-data, 2026. Hugging Face dataset, revision ea9d721, CC-BY-4.0. Ac- cessed November 2026 on Pith
- [14] Qwen Team. Qwen3-ASR toolkit (qwen-asr 0.0.6).https://github.com/QwenLM/Qwen3-ASR, 2026. PyPI package qwen-asr 0.0.6. Accessed November 2026 on Pith
- [15] A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever. Robust speech recognition via large-scale weak supervision. InProc. International Conference on Machine Learning (ICML), volume 202 ofProceedings of Machine Learning Research, pages 28492–28518. PMLR, 2023 on Pith
- [16] T. N. Sainath, R. Pang, D. Rybach, Y. He, R. Prabhavalkar, W. Li, et al. Two-pass end-to-end speech recognition. InProc. Interspeech, pages 2773–2777, 2019 on Pith
- [17] sgl-project. SGLang-Omni: High-performance multi-stage pipeline framework for omni models.https: //github.com/sgl-project/sglang-omni, 2026. GitHub repository. Accessed November 2026 on Pith
- [18] X. Shi, X. Wang, Z. Guo, Y. Wang, P. Zhang, X. Zhang, et al. Qwen3-ASR technical report, 2026. arXiv:2601.21337 on PithPith's reviewarXiv
- [19] V. Srivastav, S. Zheng, E. Bezzam, E. L. Bihan, N. R. Koluguri, P. Żelasko, S. Majumdar, A. Moumen, and S. Gandhi. Open ASR leaderboard: Towards reproducible and transparent multilingual and long-form speech recognition evaluation, 2025. arXiv:2510.06961. Leaderboard athttps:/... on PitharXiv
- [20] G. Tay, W. Ma, J. Lee, Y. Tang, D. Lee, W. Yin, et al. Back to basics: Revisiting ASR in the age of voice agents, 2026. Introduces the WildASR benchmark. arXiv:2603.25727 on PitharXiv
show 3 more
- [21] Y. Tu, Z. Peng, J. Yu, L. Dong, S. Xu, Y. Chang, et al. VibeVoice-ASR-Streaming technical report, 2026. arXiv:2609.02812 on PithPith's reviewarXiv
- [22] L. Wei, S. Zhong, S. Xu, R. Wang, R. Huang, and M. Li. SpecASR: Accelerating LLM-based automatic speech recognition via speculative decoding. InProc. 62nd ACM/IEEE Design Automation Conference (DAC), pages 1–7, 2025 on Pith
- [23] L. Zeng, H. Deng, K. Shu, and S. Wang. Self-speculative biased decoding for faster re-translation, 2025. arXiv:2509.21740. 6 A Exactness and settings Table 3:Qwen3-ASR exactness by load.Finals byte-identical to the offline reference (%), and WER (%) over each run’s requests. N... on PitharXiv
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment